Unless you're running Elasticsearch on a single node, prepare to design for failure. Designing for failure means running your cluster in multiple locations and be ready to lose a whole data center without service interruption. It's not theoretical thinking here. You WILL lose a whole data center several times during your cluster's life.
The minimum requirement for a fault tolerant cluster is:
- 3 locations to host your nodes. 2 locations to run half of your cluster, and one for the backup master node.
- 3 master nodes. You need an odd number of eligible master nodes to avoid split brains when you lose a whole data center. Put one master node in each location so you hopefully never lose the quorum.
- 2 ingest nodes, one in each primary data center.
As many data nodes as you need, split evenly between both main locations.
Architecture of a fault tolerant Elasticsearch cluster
Elasticsearch design for failure
Elasticsearch provides an interesting feature called shard allocation awareness. It allows to split the primary shards and their replica in separated zones. Allocate nodes within a same data center to a same zone to limit the odds of having your cluster go red.
cluster:
routing:
allocation:
awareness:
attributes: "rack_id"
node:
attr:
rack_id: "dontsmokecrack"Using rack_id on the ingest nodes is interesting too, as Elasticsearch will run the queries on the closest neighbours. A query sent to the ingest node located in the datacenter 1 will more likely run on the same data center data nodes.
中文总结：
-
设计容错集群的必要性：
-
在分布式环境中，故障是不可避免的，因此必须设计容错集群以应对数据中心级别的故障。
-
集群需要能够在多个地理位置运行，并确保在失去一个数据中心时服务不中断。
-
-
容错集群的最低要求：
-
3 个地理位置：两个主要数据中心运行集群的大部分节点，第三个位置用于备份主节点。
-
3 个主节点：为了避免脑裂问题，主节点数量必须为奇数，每个地理位置部署一个主节点。
-
2 个摄入节点：每个主要数据中心部署一个摄入节点。
-
多个数据节点：根据需求部署数据节点，并均匀分布在两个主要数据中心。
-
-
分片分配感知（Shard Allocation Awareness）：
-
Elasticsearch 提供了分片分配感知功能，可以将主分片和副本分片分配到不同的区域。
-
通过配置
rack_id，将同一数据中心的节点分配到同一区域，减少集群故障的风险
-
-
查询优化：
-
在摄入节点上使用
rack_id，可以使查询更可能在最近的数据中心执行，从而减少延迟。 -
示例：发送到数据中心 1 的查询更可能在数据中心 1 的数据节点上执行。
-